The Research And Implementation Of Algorithm Of Isolated Word Speech Recognition

Speech recognition is an emerging science and technology which developed in the 60s the last century and has achived great progress. Our nation's research in this field started relatively late compared to foreign countries, therefore we need to strengthen the research of speech recognition technology.The research object of speech recognition technology is speech signal processing. This technology can make the machine understand natural language of mankind and can be human-computer interaction interface. Now, most speech recognition products are implemented by software, if they can be achieved by hardware and widely used in industrial control and intelligent terminals, the meaning is very significant and the market prospect is very broad.Speech recognition includes speaker recognition, conjunction word recognition and isolated word recognition. The content of this paper is mainly applied in Chinese speech recognition chip of small-vocabulary, speaker-independent isolated word. The main work is: studying the various algorithms and optimizing the algorithms of different stages of isolated speech recognition; on the above basis, completeing the recognition system in hardware language and simulating, verifying the system. This paper researched the feature extraction, quantification, matching model algorithms of isolated word speech recognition. Among these algorithms, I select Mel Cepstral Coefficients(MFCC) as characteristic parameters, use LBG and simulated annealing algorithm to implement the quantification, choose Hidden Markov Model(HMM) as matching model. On this basis, this paper accomplished the following improvements: combining MFCC and its high-order difference together to promote the robustness of recognition; establishing codebook using only one word to improve the recognition rate; using simulated annealing algorithm in which LBG algorithm is nested to build the global optimal codebook; changing the number of states and the number of observation in HMM to make further improvement of recognition robustness.By the above improvement, the floating-point recognition rate of 32 isolated words reached 100%, and 99.2% of the fixed-point recognition rate which is the same as hardware simulation result. The robustness of recognition has been greatly improved which can still work properly even if in the case of interference.